Bounding a matching estimator: the case of a Norwegian training program

In this paper we evaluate a Norwegian vocational training rehabilitation program by comparing employment outcomes of trainees and nonparticipants using nonexperimental data. A matching estimator is used to calculate the training effect for different subgroups of the sample. We demonstrate how bounding the matching estimator can be used to evaluate the intrinsic uncertainty of estimated training effects due to selection on unobserved individual characteristics. After adjustment for observed selection into training programs we find that the overall training effect is around six percentage points. This is mainly due to a high and significant effect for individuals with a low probability of program participation. After calculating upper and lower bounds on the test-statistics used to test the hypothesis of no training effect we find that the overall effect is sensitive to unobserved selection. However, the result that the training effect is positive for individuals who are less likely to participate in a training program is not sensitive to selection bias. These individuals also have the lowest employent probabilities, which indicates potential harmful cream skimming in the Norwegian vocational rehabilitation sector.